An Automated Framework for Decomposing Memory Transactions to Exploit Partial Rollback

In this paper, we present a framework that automatically decomposes programmer-written flat transactions into closed-nested transactions. The framework relies on two key mechanisms for the decomposition. The first is a static tool that analyzes application source code and produces a compact representation of transactions´ business logic. The second is a run-time monitor that captures the actual contention level of shared objects and, relying on the outcome of the static tool, triggers the optimal closed-nested configuration for the workload at hand. We implemented this framework atop QR-CN, an open source fault-tolerant DTM written in Java. Our experimental studies conducted using the TPC-C, Vacation and Bank benchmarks reveal that the framework yields better performance than flat nesting and manual closed nesting, especially when the workload changes.